PR6540: Set the newly introduced variables ENABLE_SHARED and
SHLIBPATH_VAR in lit.site.cfg not only in the autoconf build, but also
in a cmake one.
llvm-svn: 98171
Beyond simplifying the algorithm significantly, we no longer
need to build subregion maps in RemoveDeadBindings(). This
and other changes cause a significant speedup: the time to
analyze sqlite3.c (single core) drops by 14%.
llvm-svn: 98159
directly to the maccu / maccs instructions. We handle this in
ExpandADDSUB since after type legalisation it is messy to
recognise these operations.
llvm-svn: 98150
therefore not creating ElaboratedTypes, which are still pretty-printed
with the written tag).
Most of these testcase changes were done by script, so don't feel too
sorry for my fingers.
llvm-svn: 98149
cluster analysis algorithm as RegionStore::InvalidateRegions().
Beyond simplifying the algorithm significantly, we no longer
need to build subregion maps in RemoveDeadBindings(). This
and other changes cause a significant speedup: the time to
analyze sqlite3.c (single core) drops by 14%.
llvm-svn: 98144
injected class name of a class template or class template partial specialization.
This is a non-canonical type; the canonical type is still a template
specialization type. This becomes the TypeForDecl of the pattern declaration,
which cleans up some amount of code (and complicates some other parts, but
whatever).
Fixes PR6326 and probably a few others, primarily by re-establishing a few
invariants about TypeLoc sizes.
llvm-svn: 98134
1. Be careful with cse "cheap" expressions. e.g. constant materialization. Only cse them when the common expression is local or in a direct predecessor. We don't want cse of cheap instruction causing other expressions to be spilled.
2. Watch out for the case where the expression doesn't itself uses a virtual register. e.g. lea of frame object. If the common expression itself is used by copies (common for passing addresses to function calls), don't perform the cse. Since these expressions do not use a register, it creates a live range but doesn't close any, we want to be very careful with increasing register pressure.
Note these are heuristics so machine cse doesn't make register allocator unhappy. Once we have proper live range splitting and re-materialization support in place, these should be evaluated again.
Now machine cse is almost always a win on llvm nightly tests on x86 and x86_64.
llvm-svn: 98121
Ted Kremenek